Abstract: A data-driven deep learning method using the stacked residual long short-term memory network (ResLSTM) is proposed to predict bridge seismic responses. The proposed method uses the advantage of the Long-Short Term Memory (LSTM) network for long-term sequence regression, and the residual connection structure is used to reduce the difficulty of gradient back-propagation in the deep neural network and to improve the prediction performance of the deep network with limited samples. Meanwhile, a stacked sequence structure is applied to decrease the hidden nodes of the deep neural network, further improving its prediction accuracy. Then, the proposed method is verified by numerical examples of a two-span prestressed concrete continuous girder bridge and a cable-stayed bridge with composite girder. All training and testing samples are taken from the results of incremental dynamic analysis (IDA). In addition, the proposed approach successfully predict the seismic responses of a concrete girder bridge called Meloland Overpass in USA, and the predicted responses are compared with the historical monitoring data. The results indicate that the ResLSTM network is a robust and efficient method for predicting dynamic responses of bridges under earthquake excitations with great potential in fast and accurate evaluation of seismic vulnerability of bridges.